Physics-Informed Machine Learning Reconstruction of High Resolution Ocean Subsurface Temperature Profiles From In-Situ and Satellite Observations

The irregular and incomplete coverage of in-situ ocean temperature profile observations is a major problem for various scientific applications in ocean and climate research and operational fields. However, high-resolution gridded datasets are needed to support applications. Here, we explore a physics-informed machine learning approach based on partial convolutions with multi-branch U-Net neural network structure to reconstruct the subsurface temperature profile fields with 0.1°×0.1° weekly resolution in Western Pacific Ocean. The input data include in-situ temperature profile observations, high-resolution satellite remote-sensing products (including sea surface height, sea surface temperature, sea surface salinity, etc.), and a coarse-resolution (1°× 1°) gridded subsurface temperature product (IAPv4). We show that the new reconstruction retained the large-scale features represented by the 1°× 1° temperature gridded data but added mesoscale features (because of the inputs of high-resolution satellite data). The application of physical constraints for subsurface vertical structure improves the reconstruction near thermocline. The root mean square error (RMSE) can be reduced by ~12% in the target region in average with greater improvements in the upper layer (0-700m). Further analysis shows the small-scale information is performed well also in the sparse observation coverage area with some typical mesoscale vortex features can be identified, and the features in the strait and offshore regions can be effectively improved compared with coarse resolution 1°× 1° temperature gridded data. The successful application of machine learning in this study provides confidence for the accurate reconstruction of high-resolution ocean and climate data, which can improve and complement the existing data assimilation and objective analysis methods for reconstructing multi-scale ocean information in complex regions.